prepare for collective communicate upgrade in dygraph (#54417)

wentaoyu · web-flow · commit 46c5767437ad · 2023-06-20T17:35:52.000+08:00
diff --git a/paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc b/paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc
@@ -125,7 +125,8 @@ void AsyncWorkQueue::AddTask(const OpFuncType& op_func_type,
   queue_group_->AddTask(op_func_type == OpFuncType::kGpuAsync, std::move(fn));
 }
 
-bool IsCommunicationOp(const std::string& op_name) {
+bool IsCommunicationOp(const OperatorBase* op) {
+  const std::string& op_name = op->Type();
   const std::set<std::string> special_comm_op_set = {
       "send",
       "recv",
@@ -137,14 +138,17 @@ bool IsCommunicationOp(const std::string& op_name) {
       special_comm_op_set.count(op_name)) {
     return true;
   }
+  if (op->HasAttr("ring_id")) {
+    return true;
+  }
   return false;
 }
 
 bool IsCommunicationOp(const Instruction& instr) {
   if (!instr.OpBaseValid()) {
     return false;
   }
-  return IsCommunicationOp(instr.OpBase()->Type());
+  return IsCommunicationOp(instr.OpBase());
 }
 
 bool IsCpuOp(const Instruction& instr) {
@@ -579,7 +583,7 @@ void BuildOpFuncList(const platform::Place& place,
       op_func_node.stream_priority_ = dist_attr->stream_priority();
       op_func_node.scheduling_priority_ = dist_attr->scheduling_priority();
     } else {
-      if (interpreter::IsCommunicationOp(op_type)) {
+      if (interpreter::IsCommunicationOp(op)) {
         // NOTE(Ruibiao): Dispatching computation before communication improves
         // multi-stream overlap when the time cost of communication less than
         // that of the calculation (e.g., ResNet50_bs128_pure_fp16 N4C32
diff --git a/paddle/fluid/framework/new_executor/interpreter/interpreter_util.h b/paddle/fluid/framework/new_executor/interpreter/interpreter_util.h
@@ -65,7 +65,7 @@ class AsyncWorkQueue {
   std::unique_ptr<WorkQueueGroup> queue_group_;
 };
 
-bool IsCommunicationOp(const std::string& op_name);
+bool IsCommunicationOp(const OperatorBase* op);
 
 bool IsCommunicationOp(const Instruction& instr);
 
diff --git a/paddle/fluid/operators/collective/alltoall_op.cc b/paddle/fluid/operators/collective/alltoall_op.cc
@@ -17,7 +17,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-class AllToAllOp : public framework::OperatorWithKernel {
+class AllToAllBaseOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
@@ -43,7 +43,7 @@ class AllToAllOp : public framework::OperatorWithKernel {
   }
 };
 
-class AllToAllOpMaker : public framework::OpProtoAndCheckerMaker {
+class AllToAllBaseOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   void Make() {
     AddInput("X", "(Tensor) tensor send.");
@@ -67,7 +67,9 @@ Scatter tensors from all participators to all participators.
 namespace ops = paddle::operators;
 namespace plat = paddle::platform;
 
-REGISTER_OP_WITHOUT_GRADIENT(alltoall, ops::AllToAllOp, ops::AllToAllOpMaker)
+REGISTER_OP_WITHOUT_GRADIENT(alltoall,
+                             ops::AllToAllBaseOp,
+                             ops::AllToAllBaseOpMaker)
 
 PD_REGISTER_STRUCT_KERNEL(alltoall,
                           CPU,
diff --git a/paddle/phi/core/distributed/nccl_comm_context.cc b/paddle/phi/core/distributed/nccl_comm_context.cc
@@ -19,6 +19,7 @@
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/distributed/check/nccl_dynamic_check.h"
 #include "paddle/phi/core/distributed/check/static_check.h"
+#include "paddle/phi/core/distributed/utils.h"
 #include "paddle/phi/core/enforce.h"
 #include "paddle/phi/core/utils/data_type.h"
 
@@ -61,6 +62,17 @@ void NCCLCommContext::Broadcast(phi::DenseTensor* out_tensor,
 void NCCLCommContext::AllGather(phi::DenseTensor* out_tensor,
                                 const phi::DenseTensor& in_tensor,
                                 gpuStream_t stream) {
+  phi::distributed::CommStaticCheck::GatherLikeShape(*out_tensor,
+                                                     in_tensor,
+                                                     /*dst_rank*/ rank_,
+                                                     /*cur_rank*/ rank_,
+                                                     size_);
+  if (FLAGS_enable_nccl_dynamic_check) {
+    phi::distributed::NCCLDynamicCheck::CheckShape(*out_tensor,
+                                                   /*root_rank*/ 0,
+                                                   rank_,
+                                                   nccl_comm_);
+  }
   PADDLE_ENFORCE_GPU_SUCCESS(
       phi::dynload::ncclAllGather(in_tensor.data(),
                                   out_tensor->data(),
@@ -71,29 +83,43 @@ void NCCLCommContext::AllGather(phi::DenseTensor* out_tensor,
 }
 void NCCLCommContext::ReduceScatter(phi::DenseTensor* out_tensor,
                                     const phi::DenseTensor& in_tensor,
+                                    ncclRedOp_t reduce_type,
                                     gpuStream_t stream) {
-  int64_t out_size = in_tensor.numel() / GetSize();
+  phi::distributed::CommStaticCheck::ScatterLikeShape(*out_tensor,
+                                                      in_tensor,
+                                                      /*dst_rank*/ rank_,
+                                                      /*cur_rank*/ rank_,
+                                                      size_);
+  if (FLAGS_enable_nccl_dynamic_check) {
+    phi::distributed::NCCLDynamicCheck::CheckShape(*out_tensor,
+                                                   /*root_rank*/ 0,
+                                                   rank_,
+                                                   nccl_comm_);
+  }
   PADDLE_ENFORCE_GPU_SUCCESS(
       phi::dynload::ncclReduceScatter(in_tensor.data(),
                                       out_tensor->data(),
-                                      out_size,
+                                      out_tensor->numel(),
                                       ToNCCLDataType(in_tensor.type()),
-                                      ncclSum,
+                                      reduce_type,
                                       nccl_comm_,
                                       stream));
 }
 
 void NCCLCommContext::Send(const phi::DenseTensor& in_tensor,
+                           const int64_t& count,
                            const int& peer,
                            gpuStream_t stream) {
+  phi::distributed::CommStaticCheck::CheckShape(in_tensor, rank_, size_);
+
   if (FLAGS_enable_nccl_dynamic_check) {
     NCCLDynamicCheck::CheckShape(in_tensor, rank_, rank_, nccl_comm_);
   }
 
   PADDLE_ENFORCE_GPU_SUCCESS(
       phi::dynload::ncclSend(in_tensor.data(),
-                             in_tensor.numel(),
-                             ToNCCLDataType(in_tensor.type()),
+                             count,
+                             ToNCCLDataType(in_tensor.dtype()),
                              peer,
                              nccl_comm_,
                              stream));
@@ -102,16 +128,18 @@ void NCCLCommContext::Send(const phi::DenseTensor& in_tensor,
 }
 
 void NCCLCommContext::Recv(phi::DenseTensor* out_tensor,
+                           const int64_t& count,
                            const int& peer,
                            gpuStream_t stream) {
+  phi::distributed::CommStaticCheck::CheckShape(*out_tensor, rank_, size_);
   if (FLAGS_enable_nccl_dynamic_check) {
-    NCCLDynamicCheck::CheckShape(*out_tensor, rank_, rank_, nccl_comm_);
+    NCCLDynamicCheck::CheckShape(*out_tensor, peer, rank_, nccl_comm_);
   }
 
   PADDLE_ENFORCE_GPU_SUCCESS(
       phi::dynload::ncclRecv(out_tensor->data(),
-                             out_tensor->numel(),
-                             ToNCCLDataType(out_tensor->type()),
+                             count,
+                             ToNCCLDataType(out_tensor->dtype()),
                              peer,
                              nccl_comm_,
                              stream));
@@ -123,6 +151,17 @@ void NCCLCommContext::AllReduce(phi::DenseTensor* out_tensor,
                                 const phi::DenseTensor& in_tensor,
                                 ncclRedOp_t reduce_type,
                                 gpuStream_t stream) {
+  phi::distributed::CommStaticCheck::SameShape(*out_tensor,
+                                               in_tensor,
+                                               /*dst_rank*/ rank_,
+                                               /*cur_rank*/ rank_,
+                                               size_);
+  if (FLAGS_enable_nccl_dynamic_check) {
+    phi::distributed::NCCLDynamicCheck::CheckShape(*out_tensor,
+                                                   /*root_rank*/ 0,
+                                                   rank_,
+                                                   nccl_comm_);
+  }
   PADDLE_ENFORCE_GPU_SUCCESS(
       phi::dynload::ncclAllReduce(in_tensor.data(),
                                   out_tensor->data(),
@@ -138,6 +177,17 @@ void NCCLCommContext::Reduce(phi::DenseTensor* out_tensor,
                              ncclRedOp_t reduce_type,
                              int root,
                              gpuStream_t stream) {
+  phi::distributed::CommStaticCheck::SameShape(*out_tensor,
+                                               in_tensor,
+                                               /*dst_rank*/ root,
+                                               /*cur_rank*/ rank_,
+                                               size_);
+  if (FLAGS_enable_nccl_dynamic_check) {
+    phi::distributed::NCCLDynamicCheck::CheckShape(*out_tensor,
+                                                   /*root_rank*/ root,
+                                                   rank_,
+                                                   nccl_comm_);
+  }
   PADDLE_ENFORCE_GPU_SUCCESS(
       phi::dynload::ncclReduce(in_tensor.data(),
                                out_tensor->data(),
@@ -149,5 +199,10 @@ void NCCLCommContext::Reduce(phi::DenseTensor* out_tensor,
                                stream));
 }
 
+void NCCLCommContext::GroupStart() {
+  NCCL_CHECK(phi::dynload::ncclGroupStart());
+}
+void NCCLCommContext::GroupEnd() { NCCL_CHECK(phi::dynload::ncclGroupEnd()); }
+
 }  // namespace distributed
 }  // namespace phi
diff --git a/paddle/phi/core/distributed/nccl_comm_context.h b/paddle/phi/core/distributed/nccl_comm_context.h
@@ -38,13 +38,18 @@ class NCCLCommContext final : public CommContext {
                  int root,
                  gpuStream_t stream);
   void Send(const phi::DenseTensor& in_tensor,
+            const int64_t& count,
             const int& peer,
             gpuStream_t stream);
 
-  void Recv(phi::DenseTensor* out_tensor, const int& peer, gpuStream_t stream);
+  void Recv(phi::DenseTensor* out_tensor,
+            const int64_t& count,
+            const int& peer,
+            gpuStream_t stream);
 
   void ReduceScatter(phi::DenseTensor* out_tensor,
                      const phi::DenseTensor& in_tensor,
+                     ncclRedOp_t reduce_type,
                      gpuStream_t stream);
 
   void AllGather(phi::DenseTensor* out_tensor,
@@ -62,6 +67,10 @@ class NCCLCommContext final : public CommContext {
               int root,
               gpuStream_t stream);
 
+  void GroupStart();
+
+  void GroupEnd();
+
  private:
   DISABLE_COPY_AND_ASSIGN(NCCLCommContext);
 
diff --git a/paddle/phi/core/distributed/utils.h b/paddle/phi/core/distributed/utils.h
@@ -0,0 +1,40 @@
+// Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "paddle/phi/core/dense_tensor.h"
+
+namespace phi {
+namespace distributed {
+
+inline phi::DenseTensor GetPartialTensor(const phi::DenseTensor& tensor,
+                                         int64_t offset,
+                                         int64_t numel) {
+  phi::DenseTensor tensor_flattened;
+  tensor_flattened.ShareDataWith(tensor);
+  tensor_flattened.Resize({tensor.numel()});
+  return tensor_flattened.Slice(offset, offset + numel);
+}
+
+#define NCCL_CHECK(cmd)     \
+  do {                      \
+    ncclResult_t r = cmd;   \
+    if (r != ncclSuccess) { \
+      exit(EXIT_FAILURE);   \
+    }                       \
+  } while (0)
+
+}  //  namespace distributed
+}  // namespace phi
diff --git a/paddle/phi/kernels/gpu/p_recv_kernel.cu b/paddle/phi/kernels/gpu/p_recv_kernel.cu
@@ -48,7 +48,8 @@ DDim recv_shape_info(const Context& dev_ctx,
   phi::DenseTensor* gpu_shape_size_tensor = new phi::DenseTensor(shape_dtype);
   gpu_shape_size_tensor->Resize({1});
   dev_ctx.Alloc(gpu_shape_size_tensor, shape_dtype);
-  comm_ctx->Recv(gpu_shape_size_tensor, peer, stream);
+  comm_ctx->Recv(
+      gpu_shape_size_tensor, gpu_shape_size_tensor->numel(), peer, stream);
 
   // copy the shape size tensor to cpu
   phi::DenseTensor* cpu_shape_size_tensor = new phi::DenseTensor(shape_dtype);
@@ -71,7 +72,7 @@ DDim recv_shape_info(const Context& dev_ctx,
   phi::DenseTensor* gpu_shape_tensor = new phi::DenseTensor(shape_dtype);
   gpu_shape_tensor->Resize({shape_size});
   dev_ctx.Alloc(gpu_shape_tensor, shape_dtype);
-  comm_ctx->Recv(gpu_shape_tensor, peer, stream);
+  comm_ctx->Recv(gpu_shape_tensor, gpu_shape_tensor->numel(), peer, stream);
 
   // copy the shape tensor to cpu
   phi::DenseTensor* cpu_shape_tensor = new phi::DenseTensor(shape_dtype);
@@ -141,7 +142,7 @@ void PRecvKernel(const Context& dev_ctx,
     out->Resize(new_dim);
   }
   dev_ctx.Alloc(out, dtype);
-  comm_ctx->Recv(out, peer, stream);
+  comm_ctx->Recv(out, out->numel(), peer, stream);
 #else
   PADDLE_THROW(
       errors::PreconditionNotMet("PaddlePaddle should compile with GPU."
@@ -165,7 +166,7 @@ void PRecvArrayKernel(const Context& dev_ctx,
     auto out = out_array->at(idx);
     auto out_dims = out.dims();
     dev_ctx.Alloc(&out, dtype);
-    comm_ctx->Recv(&out, peer, stream);
+    comm_ctx->Recv(&out, out.numel(), peer, stream);
     VLOG(3) << "rank " << comm_ctx->GetRank() << " recv "
             << phi::product(out_dims) << " from " << peer;
   }
diff --git a/paddle/phi/kernels/gpu/p_send_kernel.cu b/paddle/phi/kernels/gpu/p_send_kernel.cu
@@ -65,7 +65,8 @@ void send_shape_info(const Context& dev_ctx,
                      cpu_shape_size_tensor.numel() * sizeof(int),
                      stream);
 
-  comm_ctx->Send(*gpu_shape_size_tensor, peer, stream);
+  comm_ctx->Send(
+      *gpu_shape_size_tensor, gpu_shape_size_tensor->numel(), peer, stream);
   VLOG(3) << "send the shape size: " << shape_size << " to peer";
 
   // step2: send the shape
@@ -87,7 +88,7 @@ void send_shape_info(const Context& dev_ctx,
                      cpu_shape_tensor.data(),
                      cpu_shape_tensor.numel() * sizeof(int),
                      stream);
-  comm_ctx->Send(*gpu_shape_tensor, peer, stream);
+  comm_ctx->Send(*gpu_shape_tensor, gpu_shape_tensor->numel(), peer, stream);
   VLOG(3) << "send the shape: (" << dims << ") to peer";
 }
 
@@ -131,7 +132,7 @@ void PSendKernel(const Context& dev_ctx,
     send_shape_info<Context>(dev_ctx, x, comm_ctx, peer, stream);
   }
 
-  comm_ctx->Send(x, peer, stream);
+  comm_ctx->Send(x, x.numel(), peer, stream);
 #else
   PADDLE_THROW(
       errors::PreconditionNotMet("PaddlePaddle should compile with GPU."
@@ -153,7 +154,7 @@ void PSendArrayKernel(const Context& dev_ctx,
     auto x = x_array.at(idx);
     int numel = x.numel();
     ncclDataType_t dtype = ToNCCLDataType(x.type());
-    comm_ctx->Send(x, peer, stream);
+    comm_ctx->Send(x, x.numel(), peer, stream);
     VLOG(3) << "rank " << comm_ctx->GetRank() << " send "
             << phi::product(x.dims()) << " to " << peer;
   }
diff --git a/paddle/phi/kernels/gpu/reduce_scatter_kernel.cu b/paddle/phi/kernels/gpu/reduce_scatter_kernel.cu
@@ -50,7 +50,7 @@ void ReduceScatterKernel(const Context& dev_ctx,
 
   out->Resize(out_dims);
   dev_ctx.template Alloc<T>(out);
-  comm_context->ReduceScatter(out, x, stream);
+  comm_context->ReduceScatter(out, x, ncclSum, stream);
 #else
   PADDLE_THROW(
       errors::PreconditionNotMet("PaddlePaddle should compile with GPU."));
